Running the Performance Tests

Having cloned the repository, start by building, compiling and running a test in a build directory of your choice.

Example: Collisions0d Performance Test in Serial

You can build and compile this example from the root of your performance_testing_cleo directory first by building and compiling the exectuable for a build of your choice (serial, openmp, cuda or threads).

E.g. For a serial build of the colls0d example (see also cond0d, motion2d and thermo3d):

1. Build and compile:

   $ mkdir /work/bm1183/m300950/performance_testing_cleo/builds
   $ ./scripts/bash/builds.sh ./src /work/bm1183/m300950/performance_testing_cleo/builds serial openmp cuda threads
   $ ./scripts/bash/compiles.sh /work/bm1183/m300950/performance_testing_cleo/builds colls0d serial openmp cuda threads
   

Note that you might want to change the team_size in the KokkosCleoSettings namespace before compilation.

1. Setup the example:

   $ ./scripts/collisions0d/setup_colls0d.sh \
     /work/bm1183/m300950/bin/envs/perftests/bin/python \
     /home/m/m300950/CLEO \
     /home/m/m300950/performance_testing_cleo \
     /work/bm1183/m300950/performance_testing_cleo/builds \
     serial openmp cuda threads
   

2. (Or) If setting up cond0d:

   $ ./scripts/condensation0d/setup_cond0d.sh \
     /work/bm1183/m300950/bin/envs/perftests/bin/python \
     /home/m/m300950/CLEO \
     /home/m/m300950/performance_testing_cleo \
     /work/bm1183/m300950/performance_testing_cleo/builds \
     serial openmp cuda threads
   

3. (Or) If setting up motion2d:

   $ ./scripts/motion2d/setup_motion2d.sh \
     /work/bm1183/m300950/bin/envs/perftests/bin/python \
     /home/m/m300950/CLEO \
     /home/m/m300950/performance_testing_cleo \
     /work/bm1183/m300950/performance_testing_cleo/builds \
     serial openmp cuda threads
   

4. (Or) If setting up thermo3d:

   $ ./scripts/constthermo3d/setup_thermo3d.sh \
     /work/bm1183/m300950/bin/envs/perftests/bin/python \
     /home/m/m300950/CLEO \
     /home/m/m300950/performance_testing_cleo \
     /work/bm1183/m300950/performance_testing_cleo/builds \
     serial openmp cuda threads
   

#. a) Then you could perform one normal run with e.g. ./builds/serial/collisions0d/colls0d ./builds/serial/tmp/colls0d/config_1_128_1_0.yaml

1. 2. Or e.g. run the example with Kokkos performance tool(s):

   $ ./scripts/bash/run_profiling.sh \
     /work/bm1183/m300950/bin/envs/perftests/bin/python \
     /home/m/m300950/performance_testing_cleo \
     /work/bm1183/m300950/performance_testing_cleo/builds \
     colls0d \
     kerneltimer.spacetimestack.memoryevents.memoryusage \
     TRUE \
     serial openmp cuda threads
   $ ./scripts/bash/postproc_profiling.sh \
     /work/bm1183/m300950/bin/envs/perftests/bin/python \
     /home/m/m300950/performance_testing_cleo \
     /work/bm1183/m300950/performance_testing_cleo/builds \
     colls0d \
     kerneltimer.spacetimestack.memoryevents.memoryusage \
     FALSE \
     serial openmp cuda threads
   $ ./scripts/bash/create_grand_datasets.sh \
     /work/bm1183/m300950/bin/envs/perftests/bin/python \
     /home/m/m300950/performance_testing_cleo \
     /work/bm1183/m300950/performance_testing_cleo/builds \
     colls0d \
     kerneltimer \
     FALSE \
     serial openmp cuda threads
   

2. 3. You can add/remove profilers in the list of profiler names seperated by a ‘.’, e.g. kerneltimer.spacetimestack.memoryevents.memoryusage to run profilers sequentially